Align integer gen and rec loops to 64 bytes
Some checks are pending
Build / build-mac (push) Waiting to run
Build / build-linux (push) Waiting to run
Build / build-arm32 (push) Waiting to run
Build / build-win (push) Waiting to run
Build / build-freebsd (push) Waiting to run
Build / build-musl (push) Waiting to run
CodeQL Advanced / Analyze (c-cpp) (push) Waiting to run
CodeQL Advanced / Analyze (actions) (push) Waiting to run

This commit is contained in:
Andrea Mazzoleni 2026-10-05 13:46:22 +02:00
parent 143813a90e
commit 6ae014d6ed
3 changed files with 40 additions and 0 deletions

View file

@ -1055,6 +1055,12 @@ The implementation deliberately targets at most six parity levels.
Short list of ideas already tried or rejected
---------------------------------------------
* Loop alignment to 64 bytes:
improves scalar integer routines (raid/int.c) by +4% to +15% by avoiding fetch
and decode stalls on compact loops. Tried on vector SIMD routines (SSE, AVX,
NEON), with no improvement because wide unrolled loops are memory-bound and
cached in the uop cache. Kept only for integer routines.
* AVX512 128-byte generation processing:
tried, can be slower.